Abstract To improve accuracy of the prediction of soil carbon using a soil carbon calibration model, feature transformation and feature selection was done to soil infrared reflected spectral (NIRS) data obtained with a field ambulatory infrared spectroscopy device in situ. Firstly, feature transformation was done of the soil NIRS data through independent component analysis (ICA), principle component analysis (PCA) or wavelet analysis (WA), and then feature selection was through uninformative variable elimination (UVE), successive projection algorithm (SPA), uninformative variable elimination in combination with successive projection algorithm (UVE-SPA), and genetic algorithm with partial least squares regression (GA-PLS), separately. And in the end, a soil carbon calibration model was established. Results show that after the processing, a prediction model, better than subjecting the NIRS data to direct wave band selection in accuracy, can be built up, while the combination of the feature selection method with PCA or WA could only achieve some similar effects to those of subjecting NIRS data to direct wave band selection. Therefore, it is feasible to establish a more reliable soil carbon prediction model through feature transformation and selection with the feature selection method coupled with ICA of the NIRS data acquired with a field ambulatory device under complicated environmental condition.